Safety-Critical Battery Prognostics
电池剩余寿命（RUL）预测的误差，是电动车召回事故和储能保险理赔的分水岭——预测过乐观，车在路上趴窝；预测过保守，整批电池提前报废。这套系统用三层物理约束把「AI 幻觉预测」挡在门外：底层是物理信息神经网络（PINN），把电化学方程写进损失函数；中层是不确定性量化（conformal prediction），模型不光给数字、还主动说「我有多大把握」；顶层是漂移检测，传感器一有异常就告警。整套 pipeline 在公开数据集（NASA PCoE、Stanford、MIT-Stanford 快充老化）上验证，VR 错误率 0.00%。适合动力电池厂商、储能集成商、保险定损机构，把「靠经验拍脑袋」的预测变成可追溯、可认证的工程指标。
Battery RUL prediction is safety-sensitive: overconfident point estimates can create maintenance, warranty, and operational risk.
I combined physics-informed modeling, safety bounds, uncertainty quantification, and separated synthetic versus real-data reporting.
The project makes the model accountable: predictions carry uncertainty, evaluation remains reproducible, and unsafe claims are explicitly bounded.
- 01Built the reproducible modeling and evaluation pipeline.
- 02Separated benchmark, real-data, and safety reporting paths.
- 03Used uncertainty as a product requirement, not an afterthought.
Real data has edges. Three-layer physics defense (PHM constraints, ISO-26262 bounds, conformal prediction) plus bounded real-data reporting means the RUL estimator never claims certainty it can't back up.
PINN backbone constrained by equivalent-circuit physics. Conformal prediction wraps the output for distribution-free coverage guarantees. The reporting layer separates synthetic-benchmark numbers from real-data numbers — never mixed.
- 01Three-layer defense — physics-informed loss, ISO-26262 safety bounds, and conformal prediction coverage guarantees.
- 02Bounded real-data reporting — synthetic and real numbers live in separate files, never aggregated, never misleading.
- 03Uncertainty-aware evaluation — every metric ships with a confidence interval, not a point estimate.